|Publication number||US7324541 B2|
|Application number||US 10/744,198|
|Publication date||Jan 29, 2008|
|Filing date||Dec 22, 2003|
|Priority date||Dec 22, 2003|
|Also published as||US20050135355|
|Publication number||10744198, 744198, US 7324541 B2, US 7324541B2, US-B2-7324541, US7324541 B2, US7324541B2|
|Inventors||Raman Muthukrishnan, Anujan Varma|
|Original Assignee||Intel Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (37), Referenced by (23), Classifications (10), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
Store-and-forward devices, such as switches and routers, are used in packet networks, such as the Internet, for directing traffic at interconnection points. The store-and-forward devices include a plurality of line cards for receiving and transmitting data from/to external sources. The line cards are connected to one another via a backplane and a switching fabric. The backplane provides data paths between line cards and the switching fabric and the switching fabric provides configurable data paths between line cards. The line cards receiving data from external sources (ingress ports) receive data (packets) of various sizes. The data received are stored in queues prior to being transmitted to the appropriate line cards for transmission to external sources (egress ports). The packets include a header that identifies the destination of the packet. The packet is stored in the queue associated with that destination. The packet may also identify a priority for the data and the ingress port may also include queues for the various priorities.
The ingress ports send requests for transmitting data to a scheduler within the switching fabric. The scheduler generates grants for the queues that should transmit packets therefrom. The packets are switched through a crossbar switching matrix in batches. A batch consists of at most one packet selected from each input port. Thus, no more than one of the packets is destined for each output port. The packets in a batch are transferred in parallel across the crossbar switching matrix. While the packets from a scheduled batch are being transferred through the crossbar, the scheduler can select the packets to form the next batch, so that the transmission of the new batch of packets can start as soon as transmission of the current batch ends. At the end of the batch of packets, the fabric scheduler re-configures the crossbar switching matrix so as to connect the input ports to the corresponding output ports where the next packet is destined to.
Because the packets are transferred in batches, the switching paths in the crossbar switching matrix are kept unchanged for the duration of the longest packet being transferred across the crossbar in that batch. When the packets are of variable size (as is the case for packets generated by most network protocols in the industry), this results in wasted bandwidth. For example, when a 50-byte packet and a 1500-byte packet are part of the same batch, the crossbar is maintained in the same configuration for the duration of the 1500-byte packet, and only 1/30th of the bandwidth of the path is used by the 50-byte packet.
One solution for avoiding the inefficiency caused by variable-size packets is to divide the packets into fixed-size units before switching through the crossbar switching fabric, and combine the fragments into the original packet at the output of the fabric. The packet fragments switched through the crossbar are called “segments” or “cells”. The fabric scheduler selects at most one cell from each input port to form a batch, such that the destination port numbers associated with the cells in the same batch are distinct. The cells in the same batch are then transmitted in parallel. Because the cells are of the same size, no bandwidth is wasted in the crossbar. The cells switched through the fabric have a fixed size. This fixed size is typically chosen to correspond to the size of the smallest packet switched by the fabric, plus the size of any internal headers added by the router or switch before passing the packet through the fabric.
The fabric scheduler computes a new schedule for the batch of cells during the transmission time of a cell. In a high-speed switch, this time interval can be extremely short. For example, with a cell size of 64 bytes and a port rate of 10 Gigabits/second, the fabric scheduler schedules a new batch of cells every 51.2 nanoseconds. The crossbar switching matrix is also configured at intervals of 51.2 nanoseconds. As the port speed is increased, both the fabric scheduler and the crossbar reconfiguration are made correspondingly faster. This is especially a problem when an optical switching device is used as the crossbar switching matrix. While supporting very high data rates, many of the optical switching devices have long reconfiguration times. This makes them unsuitable for use in a cell-based fabric.
Another difficulty with the cell-based fabric is that it is difficult to separate the crossbar switching matrix (the data path) and the fabric scheduler, because the delays in communication between them can become a bottleneck. During every scheduling cycle, the header information (in particular, the destination port number) from cells stored in the input buffers of the crossbar matrix is passed to the fabric scheduler, and the crossbar configuration setting is communicated back from the scheduler to the crossbar matrix. If the scheduler is physically separated from the crossbar matrix (on separate chips or circuit boards), the delays in communication between the two may make it difficult to achieve the scheduling rate needed in a high-speed router or switch.
The features and advantages of the various embodiments will become apparent from the following detailed description in which:
Store-and-forward devices, such as switches and routers, are used in packet networks, such as the Internet, for directing traffic at interconnection points. Store-and-forward devices include a plurality of interface modules, a switch fabric for selectively connecting different interface modules, and a backplane for connecting the interface modules and the switching fabric. The interface modules include receivers (ingress ports) to receive data from and transmitters (egress ports) to transmit data to multiple sources (e.g., computers, other store and forward devices) over multiple communication links (e.g., twisted wire pair, fiber optic, wireless). The sources may be capable of transmitting/receiving data at different speeds, different quality of service, etc. over the different communication links. The interface modules can transmit/receive data using any number of protocols including Asynchronous Transfer Mode (ATM), Internet Protocol (IP), and (Time Division Multiplexing) TDM. The data may be variable length or fixed length blocks, such as cells, packets or frames.
The data received from external sources is stored in a plurality of queues. The queues may be stored in any type of storage device and preferably are a hardware storage device such as semiconductor memory, on-chip memory, off-chip memory, field-programmable gate arrays (FPGAs), random access memory (RAM), or a set of registers. The interface modules may be line cards or chips contained on line cards. A single line card may include a single interface module (receiver or transmitter) or multiple interface modules (receivers, transmitters, or a combination). The interface modules may be Ethernet (e.g., Gigabit, 10 Base T), ATM, Fibre channel, Synchronous Optical Network (SONET), Synchronous Digital Hierarchy (SDH) or various other types. A line card having multiple interface modules may have the same type of interface modules (e.g., ATM) or may contain some combination of different interface module types. The backplane may be electrical or optical.
The ingress fabric interface module 130 receives packets from the packet processor/traffic manager device on a line card. The packet processor/traffic manager is responsible for processing the packets arriving from the external links, determining the fabric port number associated with the incoming packet (from a header lookup), and attaching this information to the packet for use by the switching fabric 100. The ingress fabric interface module 130 receives the packets, stores the packets in associated queues, and sends the packets to the switching matrix 110 for transfer to a different line card. The egress fabric interface modules 160 are responsible for receiving packets arriving from the switching matrix 110 (typically from a different line card), and passing them on for any egress processing needed in a line card and subsequently for transmission out on the external links. It should be noted that a fabric port may aggregate traffic from more than one external port (link) associated with a line card. A pair of ingress and egress fabric interface modules 130, 160 is associated with each fabric port. When used herein the term fabric port may refer to an ingress fabric interface module and/or an egress fabric interface module. An ingress fabric interface module may be referred to as a source fabric port, a source port, an ingress fabric port, an ingress port, a fabric port, or an input port. Likewise an egress fabric interface module may be referred to as a destination fabric port, a destination port, an egress fabric port, an egress port, a fabric port, or an output port.
The ingress fabric interface modules 130 store the packets arriving from the packet processor/traffic manager in a set of queues. The packets destined to the egress fabric interface modules 160 are maintained in a separate queue (isolated from each other). In addition, the packets destined to a specific egress fabric interface module 160 can further be distributed into multiple queues based on their class of service or relative priority level. These queues may be referred to as virtual output queues. The packets may be broken down into segments and the segments stored in the queues. The segments can be variable size but are limited to a maximum size.
The segment header identifies the queue in which the segment is to be placed upon its arrival in the egress fabric interface module. The number of queues is dependent on number of priority levels (or class of services) associated with the packet. Furthermore, the number of queues may also be dependent on number of ingress fabric interface modules that can send data to the egress fabric interface module. For example, if the egress fabric interface module receives data from 8 line card ports (ingress fabric interface modules) and the line card ports support 4 levels of priority for packets to that egress fabric interface module, then the segments arriving at the egress fabric interface module may be placed in one of 32 queues (8 ingress fabric interface modules×4 priorities per ingress module). Therefore, a minimum of 5 bits are needed in the segment header to identify one of the 32 queues. The segment header also includes an “End of Packet” (EOP) bit to indicate the position of the segment within the packet where it came from. The EOP bit is set to 1 for the last segment of a packet, and 0 for the other segments. This enables the egress modules to detect the end of a packet.
The segments stored in its queues are aggregated into frames by the ingress fabric interface module 230 before transmission to the crossbar matrix 210.
The maximum size of the frame is a design parameter. The time taken to transmit the maximum-size frame is referred to as the “frame period.” This interval is the same as a scheduling interval (discussed in further detail later). The frame period can be chosen independent of the maximum packet size in the system. Typically, the frame period is chosen such that a frame can carry several maximum-size segments. The frame period is often determined by the reconfiguration time of the crossbar data path. For example, the switching time of certain optical devices are currently of the order of microseconds. If such devices are used for the data path, the frame period is on the order of microseconds. Electronic switching technologies, on the other hand, are significantly faster, allowing frame periods in the range of tens to hundreds of nanoseconds. Another factor that needs to be taken into account while choosing the frame period is the overhead in synchronizing the egress fabric interface modules with the data streams at the start of a frame. Data streams are broken at the end of a frame and the new arriving frame may be from a different ingress fabric interface module (resulting in a change in frequency and/or phase of the clock associated with the data stream). Accordingly, the egress fabric interface modules re-establish synchronization at the boundary of every frame.
The ingress fabric interface module constructs a frame by de-queuing one or more segments from its queues when instructed to do so by a grant from the fabric scheduler (discussed in further detail later). A grant may be received by an ingress fabric interface module during each frame period. The grant identifies the subset of queues from which data need to be de-queued based on the destination fabric port (egress fabric port module). This de-queuing of segments proceeds until the frame is full. Because the segments cannot further be broken up, and a frame consists of a whole number of segments, the frame constructed may not have the same size, but will be within the maximum size specified. Alternatively, the frames that do not equal the maximum frame size can be padded to the maximum size so that the frames are the same size.
By way of example, assume that the maximum frame size is 1000 bytes and that ingress port 1 just received a grant to transmit data to egress port 2 (queue 2). Assume that queue 2 has the following segments stored therein: segment 1-256 bytes, segment 2-256 bytes, segment 3-200 bytes, segment 4-256 bytes, segment 5-64 bytes, and segment 6-128 bytes. The frame would be constructed to include as many full segments as possible. In this case the first 4 segments would be selected and utilize 968 bytes. As the fifth segment cannot fit within the frame without exceeding the maximum frame size, the segment is not included. The frame is transmitted as a 968 byte frame. Alternatively, the frame can be padded to the maximum 1000 byte frame size.
If there are multiple queues (based on priority, class of service) associated with a specific destination, the ingress module chooses one or more queues from this subset based on a scheduling discipline. The scheduling discipline may be based on priority (e.g., highest priority first). That is, the queues may be serviced in order of priorities, starting from the highest priority queue and proceeding to the next priority level when the current priority level queue is empty. This de-queuing of segments proceeds through queues (priorities) until the frame is full.
By way of example, assume the same maximum frame size of 1000 bytes and that ingress port 1 just received a grant to transmit data to egress port 2 (queues 4-6 corresponding to priorities 1-3). Assume that queue 4 includes segment 1-256 bytes and segment 2-256 bytes; queue 5 includes segment 3-200 bytes; and queue 6 includes segment 4-256 bytes, segment 5-64 bytes, and segment 6-128 bytes. The frame would include segments 1 and 2 from queue 4, segment 3 from queue 5, and segment 4 from queue 6. These 4 segments selected from three different queues (priorities) generate a 968 byte frame. The frame may be transmitted as a 968 byte frame or alternatively may be padded to the maximum 1000 byte frame size.
While constructing the frame, the segments from multiple packets may be interleaved within a frame. Because the segment header provides identifying information for re-assembling them into the original packets, such interleaving does not violate data integrity. The only constraint to be satisfied is that two segments from the same packet should not be sent out of order. By way of example, assume that packet 1 includes segments 1-5 and packet 2 includes segments 6-8 and that both packets (all segments associated with) can fit within the maximum size frame. The order of the packets in the frame may be 1, 2, 3, 6, 4, 7, 8, and 5. That is the packets are interleaved within one another but the order of the segments associated with a packet are in order.
When there is only a single crossbar switching plane present, the frame is transmitted in bit-serial fashion through the crossbar planes. When multiple crossbar planes are used, the contents of the frame are striped over the available crossbar planes. Striping may be performed at the bit, byte, or word level. Additional channels may be used for protection (error detection and correction).
Referring back to
For example, assume that the ingress fabric interface module associated with fabric port 1 has five packets in its queue, two of which are destined to fabric port 3, and one each to fabric ports 5, 6 and 7, respectively. Then, the information transmitted from the ingress fabric interface module 1 in that cycle will carry at least one bit corresponding to the fabric ports 3, 5, 6 and 7 to identify the non-empty queues. The information can optionally include many other attributes, such as the amount of data in the queue and the “age” (time interval since a packet was last transmitted) of the queue. In addition, if there are multiple queues associated with each destination port, based on priority or class, then the information may include the amount of data queued at each priority level for each destination port.
A basic fabric scheduler implementation may need only the basic information (ID of non-empty queues) to be passed from the ingress fabric interface modules. More powerful scheduler implementations, supporting additional features, require more information to be passed from the ingress fabric interface modules and higher bandwidth links (or stripping of the requests over multiple links) connecting them to the scheduler.
Based on the information received from the ingress fabric interface modules 130, the fabric scheduler 120 computes a schedule for the crossbar planes 110. The schedule is computed by performing a matching of the requests received from the ingress fabric interface modules 130 and resolving any conflicts therebetween. For example, assume ingress fabric interface module 1 has packets queued for destinations 5 and 7, while ingress fabric interface module 2 has packets queued for destinations 5 and 9. During the matching phase, the scheduler 120 could match both of the ingress modules to destination 5. However, the scheduler 120 would realize the conflict and modify the schedule accordingly. The scheduler 120 may schedule ingress module 1 to send packets to destination 7 and ingress module 2 to send to destination 5, enabling both transmissions to occur in parallel during the same frame cycle. In practice, the fabric scheduler 120 may use criteria such as, the amount of data queued at various priorities, the need to maintain bandwidth guarantees, and the waiting times of packets, in the computation of the schedule.
The scheduler 120 then sets the crossbar matrix (planes) 110 to correspond to this setting. For example, if the fabric scheduler 120 generates a schedule in which the ingress fabric interface module 1 is to transmit a packet in its queue to destination port 4 in the current cycle, the scheduler 120 configures the crossbar matrix 110 to connect ingress port 1 to egress port 4 during the current frame cycle. If there are multiple crossbar planes used to stripe the data, then the planes are set in parallel to the same configuration.
After the fabric schedule 120 computes its schedule, the scheduler 120 communicates back to the ingress fabric interface modules 130 the schedule information (grants) computed. The information sent to particular ingress modules includes, at a minimum, the destination fabric port number to which it was matched. Upon receiving this information, the ingress fabric interface modules 130 de-queue data (segments) from the associated queue(s) and transmit the data (frames) to the crossbar data planes (previously discussed). This is done in parallel by the interface modules 130. Because the fabric scheduler 120 sets the crossbar planes 110 to correspond to the schedule information (grants) communicated to the ingress fabric interface modules 130, the data transmitted by the ingress modules 130 will reach the intended destination egress interface modules 160.
While communicating the schedule information (grants) to the ingress fabric interface modules 130, the fabric scheduler 120 may optionally send information about the computed schedule to the egress fabric interface modules 160. Specifically, the scheduler 120 may send to the egress modules 160 the port number associated with the ingress modules 130 that will be transmitting data to it in that cycle. Although this information can be provided within the data stream itself (as part of header), sending it directly from the fabric scheduler 120 enables the egress modules 160 to detect errors by comparing the source of the arriving data (obtained from the headers) with the scheduler-supplied port number. A mismatch indicates an error or failure in the switch fabric system. The arriving data can be discarded in such an event, thus avoiding delivery of data to an unintended port.
The crossbar switching planes 420 and the fabric scheduler 430 reside on one or more switch cards. The backplane 440 (serial channels) form the data path over which packets are transported through the crossbar switching planes 420. When the bandwidth of a single serial channel (link) is inadequate to support the data rate of the fabric port, data is striped over multiple channels. Such striping can be at different granularities (e.g., bit, byte, word). If the data is stripped over several channels, there will be a corresponding number of crossbar planes. The crossbar planes may be separate crossbar matrixes or may be a single crossbar matrix containing multiple planes. Additionally, more links and switching planes may be used to provide speedup, redundancy, error detection and/or error recovery.
The scheduling requests sent from the ingress fabric interface module to the fabric scheduler during each frame period may be formatted as request frames.
The major part of the request frame 600 is the set of requests 630, one for each destination fabric port and priority level. Assuming an example system with 64 fabric ports and 4 priority levels, there would be 256 (64 ports×4 priorities/port) distinct requests 630 in the request frame 600. The request 630 indicates that there is data in an associated queue available for transmission. The request 630 may summarize the amount of data in the associated queue. The length of the requests 630 (e.g., number of bits) may be chosen taking into account limitations on the total length of the request frame 600, and the granularity of the amount of data in the associated queue needed by the scheduler (scheduling algorithms). For example, the request 630 may be encoded as 4 bits, thus providing 16 different options for defining the amount of data in the queue. That is, the request 630 can utilize 4 bits to describe the amount of data in the queue. The requests 630 can be encoded in various ways to define the amount of data in the associated queue. One approach is to quantize the amount of data it in terms of the frame period. That is, the request 630 may be encoded to indicate the number of data frames it would take to transport the data within the associated queue over the crossbar planes.
In making scheduling decisions, the scheduler may base the decision on the priority of data. For example, if the request frame indicates that ingress module 1 priority 1 has 0.25 frame queued for egress module 3, and ingress module 2 priority 2 has 1.00 frame queued for egress module 3, then the fabric scheduler will chose the ingress module with the higher priority (ingress module 1) in making scheduling decisions for which ingress module should transmit data to egress module 3. In order to maintain high throughput, the scheduler also gives preference to requests to queues having full frames worth of data to send. For example, if the request frame indicates that ingress module 1 has only 0.25 frame of priority 1 queued for egress module 7, while ingress module 2 has 0.5 frame of priority 1 data queued for egress module 7, the fabric scheduler will select the ingress module having more data queued (ingress module 2) to transmit data to egress module 7. When the amount of data for a specific egress module and priority is equal, the scheduler may look to the total amount of data queued for the egress module. For example, if the request frame indicates that ingress module 1 has only 0.25 frame of priority 1 queued for egress module 9, and that ingress module 2 has 0.25 frame of priority 1 and 1.00 frame of priority 2 queued for egress module 9, then the fabric scheduler may select the ingress module having more data queued in total for egress module 9 (ingress module 2) as the amount of data for the highest priority was equal.
Because the scheduler may use a number of external criteria in making scheduling decisions, such as the amount of data queued at each priority level, the relative priorities of the requests, and the efficiency of the data path (how full the data frames will be), the scheduling process can be extremely difficult. According to one embodiment, the external set of scheduling criteria (priority, amount) is mapped into a set of internal scheduler priority levels (SPL).
The scheduler utilizes the SPLs (8 distinct priority levels) instead of external criteria to make scheduling decisions. In implementation, this mapping can be performed via a programmable lookup table that can be set based on the desired performance objectives. It should be noted that the above example only differentiated between full and partial frames. If the SPL was 4 bits and could thus have 16 distinct priorities, then the SPLs could distinguish between up to four different volumes of frames (e.g., 0 to ¼ frame, ¼ to ½ frame, ½ to ¾ frame, ¾ to 1; <1 frame, 1 to 2 frame, 2 to 3 frame, >3 frame).
Low priority partial frames will have low SPLs and data within the associated queues may never get scheduled (those requests may be starved). To prevent starvation of these requests (and the data within those queues) “age timers” will be implemented. The scheduler maintains an age timer for the request so that SPLs for requests that are old can be increased. For each ingress module, the scheduler maintains an N×P age timer value table (where N is the number of egress ports in the system and P is number of external priority levels). The age timer for a particular queue is incremented if the length is non-zero at the end of a scheduling cycle. The timer is reset to zero if data was transmitted in the previous frame period to the egress port and priority level associated with the age timer. A grant-service-bitmap (described later), transmitted from the ingress modules to the fabric scheduler as part of the request frames, indicates that data from a particular priority queue was sent as part of the previous data frame.
According to one embodiment, the age timers may be used to indicate when a queue has aged (exceeded some age timer threshold). A maximum age value (threshold) may be set lower for lower priorities. That is, the threshold for a priority 0 queue may be 30, while the threshold for a priority 3 may be 120. Alternatively, the age timers may provide an aging status that represents the amount of time on a continuum that the queue has not been processed. For example, if the queue has not been serviced for less than 30 cycles the aging status is a 0, if it is between 30-60 cycles it is a 1, between 60-90 cycles a 2 and so on. The higher the aging status the higher the priority that the queue is given by the scheduler. The aging status may be utilized in the formation of the SPLs.
The arbitration block 920 generates a switching schedule (ingress port to egress port links) based on the contents of the queues and the SPLs assigned thereto. The arbitration block 920 includes arbitration request blocks 930, grant arbiters 940 and accept arbiters 950. The arbitration request blocks 930 are associated with specific ingress modules. The arbitration request block 930 generates requests (activates associated bit) for those queues having requests for data to be transmitted. The arbitration request block 930 sends the requests one SPL at a time. That is, the arbitration request block 930 will send the SPL 0 requests when the arbitration block is set to SPL 0 (discussed in more detail later). For example, if ingress port 0 has SPL 0 data to be transmitted for egress ports 7, 11 and 21 it will send the corresponding requests to the grant arbitrators 940 associated with egress ports 7, 11 and 21. As the arbitration request blocks 930 are sending the requests for the same SPL (and thus same priority) there is no need for the requests to include the priority of the data. In fact, according to one embodiment the requests will be nothing more than an active bit (e.g., set to 1) to indicate that there is a request by the associated queue to transmit data. The active bit request is known as a request valid.
The grant arbiters 940 are associated with specific egress modules. The grant arbiters 940 are coupled to the arbitration request blocks 930 and are capable of receiving requests (request valids) from any number (0-to 63 in this example) of the arbitration request blocks 930. If a grant arbiter 940 receives multiple requests, the grant arbiter 940 will accept one of the requests (grant) based on some type of arbitration (e.g., round robin). For example, if ingress module 17 was the last request to be granted for egress module 11, and egress module 1 receives requests from ingress modules 10, 16, 19 and 21, the request for ingress module 19 will be granted as it is the next ingress module after the last processed ingress module. A grant valid corresponding to the request granted will be activated (e.g., set to 1) for the associated ingress port. The accept arbiters 950 are associated with specific ingress modules. The accept arbiters 950 are coupled to the grant arbiters 940 and are capable of receiving grants (grant valids) from any number (0 to 63 in this example) of the grant arbiters 940. If an accept arbiter 950 receives multiple grants, the accept arbiter 950 will accept one of the grants based on some type of arbitration (e.g., round robin). For example, if egress module 19 was the last grant to be accepted for ingress module 1, and ingress module 1 receives grants for egress modules 16, 18 and 21, the grant for egress module 21 will be accepted as it is the next egress module after the last processed egress module. An accept valid corresponding to the grant accepted will be activated (e.g., set to 1) for the associated egress port. If any accept arbiter 950 accepts a grant, the arbitration request block 930 associated with that ingress port is disabled for the remainder of the scheduling cycle. Likewise, if a grant is accepted for a certain grant arbiter 940, the grant arbiter is disabled for the remainder of the scheduling cycle.
Iterations of the scheduling process consists of the three phases: requests generated, requests granted, and grants accepted. At the end of an iteration the process continues for ingress and egress ports that were not previously associated with an accepted grant. The scheduler also includes an iteration manager (not illustrated in
In addition, unused iterations from one SPL are carried over to the next SPL. That is, if three iterations are dedicated for requests at SPL 0, but in iteration 1 it is determined that there are no requests at SPL 0, then the remaining two iterations could be used by requests at SPL 1. Thus, SPL 1 could utilize up to four iterations to grant valid requests for as many ingress to egress matches as possible. This iterative matching scheme (permitting rollover of SPLs) is referred to as Iterations Carried Over Next priority (ICON). The programmed value (maximum cumulative iteration) from
The iteration manager can increment the SPL if (1) the number of iterations dedicated for that SPL is completed, or (2) the grant arbiters indicate that no new grants can be given (either because the grant arbiters did not receive any requests, or they are already matched). It should be noted that it will take at least one iteration to determine that no new grants can be issued (no matching can be done).
Referring back to
After the iterations have been completed, the scheduler performs a pointer update operation during which the round-robin pointers in the grant arbiters 940 and the accept arbiters 950 are updated as follows: If the request pointed by the round-robin pointer for a specific SPL in a grant arbiter 940 was matched successfully, (that is, the request was satisfied on completion of the iterations), then the round-robin pointer is updated to the next valid request (in round-robin order), considering the valid requests at that SPL and higher SPLs. In the above example, if the request from ingress port 12 was satisfied, the round-robin pointer 0 in the grant arbiter will be updated to 21, the ingress port number corresponding to the next valid request at SPL-0. If, on the other hand, the grant to ingress port 12 was not accepted, the round-robin pointer for SPL-0 in the grant arbiter is not modified (so that the grant arbiter will again select the request from ingress port 12 in the next scheduling cycle, during an iteration at SPL-0).
In a more complex example, assume that the round-robin pointer for SPL-1 in a grant arbiter 0 is pointing to port 12, and that the current iteration is being performed at the SPL-1 priority level. Assume that there are active SPL-1 requests for egress port 0 from ingress ports 5, 12, 17 and 21 during the current iteration; and that there are unmatched SPL-0 requests from ingress ports 6, 16, 19 and 25, which are also submitted during the current iteration. Also assume that the scheduler was able to satisfy the SPL-1 request from port 12. Then during the pointer update operation, the round-robin pointer for SPL-1 is updated to 16, the port number corresponding to the next valid request, considering both SPL-0 and SPL-1 requests submitted during the iteration.
The round-robin pointers in the accept arbiters 950 are updated during the pointer update operation in a similar manner, except that, while updating a round-robin pointer associated with a port i, only the requests originating from the ingress port i are considered. Thus, if the request pointed by the round-robin pointer for a specific SPL in an accept arbiter i was matched successfully, then the round-robin pointer is updated to the next valid request (in round-robin order), considering the valid requests from ingress port i at that SPL and higher SPLs. For example, assume that there are active SPL-0 requests from ingress port 0 for egress ports 5, 12, 17 and 21; and active SPL-1 requests for egress ports 6, 15, 19 and 25. Also assume that the current iteration is at SPL-1, during which the scheduler was able to satisfy the SPL-1 request to egress port 15. Then during the pointer update operation, the round-robin pointer for SPL-1 is updated to 17, the port number corresponding to the next valid request, considering both SPL-0 and SPL-1 requests.
In one embodiment, the pointer updates will be performed considering the request that was satisfied was for a full frame of data or less than a frame at a single priority level. If the request that was satisfied was for less than a full frame of data, then the lower-priority requests for the same pair of ingress-egress ports may also be satisfied by the same grant, so that the pointers for the corresponding SPLs can be updated as if their requests were satisfied. The arbitration request block 930 uses the SPLs in two different ways depending on whether the scheduler is arbitrating (scheduling phase) or updating the pointers (pointer update phase). During the scheduling phase, the arbitration request block 930 sends out the requests for the SPL at which the arbitration is currently processing, as well as the requests remaining unmatched at all lower SPLs (higher priorities). For example, if SPL 2 is the current level of the iteration, then the unmatched SPL 0, SPL 1, SPL 2 requests are sent out for arbitration. During the pointer update phase, the grant and accept arbiters update their per-SPL round-robin pointers one SPL at a time. Therefore, the arbitration request block 930 sends out only requests of the SPL that is currently being processed.
The priority encoder 1110 further includes an SPL input 1170, an update pointers input 1180 and an enable scheduling input 1190. The inputs are received from the iteration manager. The SPL input 1170 receives an input SPL corresponding to the SPL of the current iteration and the priority encoder 1110 performs scheduling based on the input SPL. The enable scheduling input 1190 becomes active when the scheduler (grant arbiter 1100) is in the scheduling phase (is arbitrating). The update pointers input 1180 becomes active when the scheduler (grant arbiter 1100) is in the pointer update phase (updating the pointers). The priority encoder 1110 further includes a pointer input 1195 for receiving the round robin pointers 1120 for each SPL. The round robin pointers 1120 indicate the starting ingress port to have a request granted. Scheduling will begin at this ingress port, that is, the priority encoder will select the first active request valid signal starting at this position and proceeding cyclically through the requests.
When the enable scheduling signal is active (enable scheduling input 1190 receives an active input), the priority encoder 1110 utilizes the input SPL 1170 and the round robin pointer 1120 for the input SPL to select a request valid signal to be granted. The grant valid signal corresponding to the selected request is then made active. The grant valid signals are sent to the accept arbiters.
When the update pointers signal 1180 is active, the grant arbiter 1100 updates the round-robin pointer value corresponding to the input SPL, as previously described. Initially, the pointer value corresponding to the input SPL is read. A determination of the next valid request after the pointer value is made by utilizing the priority encoder (Find Next One logic) 1110. If the request valid signal corresponding to the current pointer value is inactive (indicating that the request was matched in the current cycle), the pointer is updated with the output of the priority encoder. Otherwise, the pointer value remains unchanged. The request blocks mask out the requests matched during the current scheduling cycle while the pointers are getting updated. Thus, in the case where the request pointed to by the current pointer value has been matched in the current frame, during the pointer update phase the request will be seen inactive by the grant arbiter block, and hence the pointer will be updated to the next valid request.
The accept arbiter 1200 further includes an SPL input 1270, an update pointers input 1280 and an enable scheduling input 1290. These inputs are received from the iteration manager. The SPL input 1270 receives an input SPL (current SPL) and the priority encoder 1210 performs scheduling based on the input SPL. The enable scheduling input 1290 becomes active when the scheduler (accept arbiter 1200) is in the scheduling phase (is arbitrating). The update pointers input 1280 becomes active when the scheduler (accept arbiter 1200) is in the pointer update phase (updating the pointers). The priority encoder 1210 further includes a pointer input 1295 for receiving the round robin pointers 1220 for each SPL. The round robin pointers 1220 indicate the starting egress port to have a request granted. Scheduling will begin at this egress port, that is, the priority encoder will select the first active grant valid signal starting at this position and proceeding cyclically through the grants.
When the enable scheduling signal 1290 is active, the priority encoder 1210 utilizes the input SPL and the round robin pointer for the input SPL to select a grant valid signal to be accepted. The accept valid signal corresponding to the selected grant is then made active. Once an accept valid is activated, the accept arbiter 1200 sends the address 1298 of the selected egress module to the request pre-processing block, which then forms the grant message.
When the update pointers signal 1280 is active, the accept arbiter 1200 updates the round-robin pointer value corresponding to the input SPL, as previously described. Initially, the pointer value corresponding to the input SPL is read. A determination of the next valid request after the pointer value is made by utilizing the priority encoder (Find Next One logic) 1210. The inputs to the priority encoder now are the requests from the ingress port corresponding to the accept arbiter. If the accept valid signal corresponding to the current pointer value is inactive (indicating that the grant was accepted in the current cycle), the pointer is updated with the output of the priority encoder. Otherwise, the pointer value remains unchanged. The request blocks mask out the requests matched during the current scheduling cycle while the pointers are getting updated. Thus, in the case where the request pointed to by the current pointer value has been accepted in the current frame, during the pointer update phase the request will be seen inactive by the accept arbiter block, and hence the pointer will be updated to the next valid accept.
As previously mentioned, a grant is formed by the scheduler at the end of the scheduling cycle for transmission to the ingress port and the egress ports. The grant is sent by the fabric scheduler in a grant frame similar to the request frame it receives from the ingress fabric interface modules. The grant to the ingress module identifies which egress module it should send data to. The ingress module knows which queues are associated with the egress module and thus can form the data frame from those queues. According to one embodiment, the segments used to create the data frame are selected from the highest priority queues first (e.g., priority 0). If the highest priority queue does not have enough segments or has no data, the ingress module moves to the next priority queue to generate the frame. Once the frame is full, and the crossbar is configured, the frame is transmitted to the switching fabric. Starting with the highest priority queues may cause some queues to starve (have segments that are not sent for long periods of time). As previously discussed, the scheduler may take account of the age of the data within the queues in the formation of the SPLs (e.g., partial aged frame is given same SPL as full frame) and thus the selection of requests for that arbitration request block (ingress module). However, if the scheduler selects a request from a particular ingress module for a particular egress module based on a lower level priority, the particular ingress module may have received higher level data prior to the dequeuing of data from queues associated with the particular egress port. Accordingly, the higher priority queue will be used to form the frames and the lower priority queue may again not be serviced.
According to one embodiment, the grant for the ingress module may include the priority as well as the egress module. When the priority level is identified, the ingress module will start dequeuing data from the identified priority queue first. Starting from a specific priority level would prevent the lower priority queues from starving (not being transmitted for a long time). The priority may be included all of the time, or only when the age timers indicate that the data in the queue is starving or is soon to be starving (e.g., has exceeded some time threshold).
The grant message to the ingress module is sent out at the end of the scheduling/grant frame (stage II), so as to give the scheduler enough time to compute the matching. That is, the scheduler can start transmitting the grant frame while it is still working on the computation of the matching, and the result is needed only close to the end of the frame time. This avoids an additional frame time worth of delay in the scheduling pipeline.
The grant frame to the egress module identifies the ingress module that will be transmitting data to the egress module. The egress module can compare the frame received from the crossbar with the grant to determine if the correct data was sent. A mismatch in the frame received and the grant indicates some type of error in transmission and results in discarding of the frame.
If both the ingress and egress modules associated with the same fabric port are packaged together (e.g., in the same chip or board), the two grant messages could be combined into a single grant frame.
If the grant includes the priority, the ingress module should start de-queuing data at the given priority. However, data from other priority levels may be sent as part of the same frame if the specified priority level does not have a full frame of data. It is important for the fabric scheduler to receive information from the ingress module on which priorities were actually serviced while sending the frame. This information is sent back from the ingress module to the fabric scheduler as a grant service bitmap.
Although the various embodiments have been illustrated by reference to specific embodiments, it will be apparent that various changes and modifications may be made. Reference to “one embodiment” or “an embodiment” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment. Thus, the appearances of the phrase “in one embodiment” appearing in various places throughout the specification are not necessarily all referring to the same embodiment.
Different implementations may feature different combinations of hardware, firmware, and/or software. For example, some implementations feature computer program products disposed on computer readable mediums. The programs include instructions for causing processors to perform techniques described above.
The various embodiments are intended to be protected broadly within the spirit and scope of the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4092732||May 31, 1977||May 30, 1978||International Business Machines Corporation||System for recovering data stored in failed memory unit|
|US4331956||Sep 29, 1980||May 25, 1982||Lovelace Alan M Administrator||Control means for a solid state crossbar switch|
|US4335458||Apr 27, 1979||Jun 15, 1982||U.S. Philips Corporation||Memory incorporating error detection and correction|
|US4695999||Sep 2, 1986||Sep 22, 1987||International Business Machines Corporation||Cross-point switch of multiple autonomous planes|
|US5127000||Aug 9, 1990||Jun 30, 1992||Alcatel N.V.||Resequencing system for a switching node|
|US5191578||Jun 14, 1990||Mar 2, 1993||Bell Communications Research, Inc.||Packet parallel interconnection network|
|US5260935||Mar 1, 1991||Nov 9, 1993||Washington University||Data packet resequencer for a high speed data switch|
|US5274758||Dec 28, 1992||Dec 28, 1993||International Business Machines||Computer-based, audio/visual creation and presentation system and method|
|US5442752||Jan 24, 1992||Aug 15, 1995||International Business Machines Corporation||Data storage method for DASD arrays using striping based on file length|
|US5483523||Aug 10, 1994||Jan 9, 1996||Alcatel N.V.||Resequencing system|
|US5535221 *||Aug 13, 1992||Jul 9, 1996||Fujitsu Limited||Frame length control in data transmission using ATM network|
|US5649157||Mar 30, 1995||Jul 15, 1997||Hewlett-Packard Co.||Memory controller with priority queues|
|US5682493||Aug 26, 1995||Oct 28, 1997||Sun Microsystems, Inc.||Scoreboard table for a counterflow pipeline processor with instruction packages and result packages|
|US5832278||Feb 26, 1997||Nov 3, 1998||Advanced Micro Devices, Inc.||Cascaded round robin request selection method and apparatus|
|US5848434||Dec 9, 1996||Dec 8, 1998||Intel Corporation||Method and apparatus for caching state information within a directory-based coherency memory system|
|US5859835||Apr 15, 1996||Jan 12, 1999||The Regents Of The University Of California||Traffic scheduling system and method for packet-switched networks|
|US5860097||Sep 23, 1996||Jan 12, 1999||Hewlett-Packard Company||Associative cache memory with improved hit time|
|US5898688 *||May 24, 1996||Apr 27, 1999||Cisco Technology, Inc.||ATM switch with integrated system bus|
|US5978951||Sep 11, 1997||Nov 2, 1999||3Com Corporation||High speed cache management unit for use in a bridge/router|
|US6055625||Jul 7, 1997||Apr 25, 2000||Fujitsu Limited||Pipeline computer with a scoreboard control circuit to prevent interference between registers|
|US6061345||Sep 30, 1997||May 9, 2000||Electronics And Telecommunications Research Institute||Crossbar routing switch for a hierarchical crossbar interconnection network|
|US6167508||Jun 2, 1998||Dec 26, 2000||Compaq Computer Corporation||Register scoreboard logic with register read availability signal to reduce instruction issue arbitration latency|
|US6170032||Dec 15, 1997||Jan 2, 2001||Texas Instruments Incorporated||Priority encoder circuit|
|US6188698||Jan 26, 1998||Feb 13, 2001||Cisco Technology, Inc.||Multiple-criteria queueing and transmission scheduling system for multimedia networks|
|US6282686||Sep 24, 1998||Aug 28, 2001||Sun Microsystems, Inc.||Technique for sharing parity over multiple single-error correcting code words|
|US6321306||Nov 9, 1999||Nov 20, 2001||International Business Machines Corporation||High performance multiprocessor system with modified-unsolicited cache state|
|US6359891||Jun 13, 2000||Mar 19, 2002||Conexant Systems, Inc.||Asynchronous transfer mode cell processing system with scoreboard scheduling|
|US6408378||May 31, 2001||Jun 18, 2002||Intel Corporation||Multi-bit scoreboarding to handle write-after-write hazards and eliminate bypass comparators|
|US7161906 *||Dec 14, 2001||Jan 9, 2007||Agere Systems Inc.||Three-stage switch fabric with input device features|
|US7212525 *||Jan 16, 2002||May 1, 2007||Hitachi, Ltd.||Packet communication system|
|US7224703 *||Dec 12, 2001||May 29, 2007||Telefonaktiebolaget Lm Ericsson (Publ)||Method and apparatus for segmenting a data packet|
|US20010009552 *||Feb 9, 2001||Jul 26, 2001||Coree1 Microsystems, Inc.||Scheduling techniques for data cells in a data switch|
|US20010038629 *||Mar 29, 2001||Nov 8, 2001||Masayuki Shinohara||Arbiter circuit and method of carrying out arbitration|
|US20030182480 *||Mar 19, 2003||Sep 25, 2003||Anujan Varma||Selecting a queue for service in a queuing system|
|US20040017778 *||Mar 24, 2003||Jan 29, 2004||Akash Bansal||Error detection and recovery of data in striped channels|
|US20040037302 *||Mar 24, 2003||Feb 26, 2004||Anujan Varma||Queuing and de-queuing of data with a status cache|
|US20050015388 *||Jul 18, 2003||Jan 20, 2005||Subhajit Dasgupta||Maintaining aggregate data counts for flow controllable queues|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7570654||Dec 22, 2003||Aug 4, 2009||Intel Corporation||Switching device utilizing requests indicating cumulative amount of data|
|US7623524||Dec 22, 2003||Nov 24, 2009||Intel Corporation||Scheduling system utilizing pointer perturbation mechanism to improve efficiency|
|US7840873||Dec 13, 2006||Nov 23, 2010||Globalfoundries Inc.||Partial CRC insertion in data packets for early forwarding|
|US7881303 *||Dec 13, 2006||Feb 1, 2011||GlobalFoundries, Inc.||Command packet packing to mitigate CRC overhead|
|US8175085 *||Jan 14, 2009||May 8, 2012||Fusion-Io, Inc.||Bus scaling device|
|US8307147 *||Sep 9, 2005||Nov 6, 2012||Freescale Semiconductor, Inc.||Interconnect and a method for designing an interconnect|
|US8572296||Jun 30, 2005||Oct 29, 2013||Freescale Semiconductor, Inc.||Device and method for arbitrating between direct memory access task requests|
|US8625624 *||Jun 13, 2008||Jan 7, 2014||Cisco Technology, Inc.||Self-adjusting load balancing among multiple fabric ports|
|US20050135356 *||Dec 22, 2003||Jun 23, 2005||Raman Muthukrishnan||Switching device utilizing requests indicating cumulative amount of data|
|US20050135398 *||Dec 22, 2003||Jun 23, 2005||Raman Muthukrishnan||Scheduling system utilizing pointer perturbation mechanism to improve efficiency|
|US20050207436 *||Mar 18, 2004||Sep 22, 2005||Anujan Varma||Switching device based on aggregation of packets|
|US20080148131 *||Dec 13, 2006||Jun 19, 2008||Hughes William A||Command Packet Packing to Mitigate CRC Overhead|
|US20080148135 *||Dec 13, 2006||Jun 19, 2008||Hughes William A||Partial CRC Insertion in Data Packets for Early Forwarding|
|US20080159145 *||Dec 29, 2006||Jul 3, 2008||Raman Muthukrishnan||Weighted bandwidth switching device|
|US20080307145 *||Sep 9, 2005||Dec 11, 2008||Freescale Semiconductor, Inc.||Interconnect and a Method for Designing an Interconnect|
|US20090125647 *||Jun 30, 2005||May 14, 2009||Citibank, N.A.||Device And Method For Executing A DMA Task|
|US20090144589 *||Jun 30, 2005||Jun 4, 2009||Freescale Semiconductor, Inc.||Device and method for controlling an execution of a dma task|
|US20090216917 *||Jun 30, 2005||Aug 27, 2009||Freescale Semiconductor, Inc||Device and method for arbitrating between direct memory access task requests|
|US20090219913 *||Feb 22, 2007||Sep 3, 2009||Kyocera Corporation||Communication Device and Method of Controlling Communication Device|
|US20090245245 *||Jan 14, 2009||Oct 1, 2009||Aprius Inc.||Bus Scaling Device|
|US20090327563 *||Jun 30, 2008||Dec 31, 2009||Ilia Greenblat||Connecting between data-handling circuits of an integrated circuit|
|US20170099369 *||Oct 1, 2015||Apr 6, 2017||Oracle International Corporation||Method and Apparatus for Conversion between Logical and Physical Frames For a Serial Link Application|
|USRE44487 *||Sep 22, 2011||Sep 10, 2013||Globalfoundries Inc.||Command packet packing to mitigate CRC overhead|
|U.S. Classification||370/429, 370/462|
|International Classification||H04L12/54, H04L12/56|
|Cooperative Classification||H04L49/9094, H04L49/254, H04L47/50|
|European Classification||H04L12/56K, H04L49/25E1, H04L49/90S|
|Aug 9, 2004||AS||Assignment|
Owner name: INTEL CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:MUTHUKRISHNAN, RAMAN;VARMA, ANUJAN;REEL/FRAME:015662/0940;SIGNING DATES FROM 20040519 TO 20040614
|Jul 21, 2011||FPAY||Fee payment|
Year of fee payment: 4
|Jul 15, 2015||FPAY||Fee payment|
Year of fee payment: 8